A deep learning-based regional rainfall spatial distribution state restoration method

Abstract

This application discloses a method for reconstructing the spatial distribution of regional rainfall based on deep learning, belonging to the field of deep learning technology. The method includes: acquiring the elevation and monthly rainfall of each grid within a target area; dividing the area into feature sub-regions based on the numerical distribution of the grid's elevation, slope, monthly rainfall, and monthly rainfall variation gradient; randomly deploying rain gauges within each feature sub-region to construct a feature mask; selecting rainfall reference data to correct rainfall observation data to obtain rainfall label data; selecting rainfall label data and feature masks from the training period to train the model; selecting rainfall label data and feature masks from the testing period to test the trained model; and obtaining recommended deployment areas for rain gauges based on the test results. This application can more accurately reconstruct the spatial distribution of regional rainfall under limited rain gauge conditions, effectively supporting extreme weather prediction and water resource assessment, while optimizing the station network layout and achieving a balance between accuracy improvement and cost control.

Description

Technical Field

[0001] This application relates to the field of deep learning technology, and in particular to a method for reconstructing the spatial distribution of regional rainfall based on deep learning. Background Technology

[0002] Regional rainfall spatial distribution refers to the spatial variation pattern and structural characteristics of rainfall at different geographical locations within a specific time scale. Obtaining accurate regional rainfall spatial distribution can reflect the distribution of rainfall intensity, rainfall center location, and rainfall range at the regional scale, which is of great significance for water resource assessment and disaster prevention under climate change.

[0003] Currently, related technologies can use traditional spatial interpolation methods such as inverse distance weighted interpolation and Kriging interpolation to reconstruct the spatial distribution of rainfall based on rain gauge observation data. However, this method still has considerable uncertainty, mainly in the following two aspects.

[0004] On the one hand, the limited number and uneven distribution of rain gauges mean that only limited, discrete rainfall information can be obtained, leading to insufficient sampling of regional rainfall data. Combined with the spatial non-uniformity and continuity of actual rainfall events, relying solely on single-point observations makes it difficult to accurately characterize the overall features of regional rainfall. Furthermore, interpolation calculations primarily rely on the spatial distance relationships and statistical correlations between rain gauges to estimate rainfall at unknown locations, failing to utilize the region's topography and macroscopic rainfall characteristics, thus hindering the improvement of the accuracy of reconstructing the spatial distribution of regional rainfall.

[0005] On the other hand, under the same station density, different station locations imply different regional topographic information and macroscopic rainfall characteristics, resulting in differences in the ability of the rain gauge network to characterize the spatial structure of regional rainfall. When selecting rain gauge stations, if the regional topographic conditions and rainfall distribution characteristics are not fully considered during station site selection, the observation data provided by improperly located rain gauge stations may fail to effectively reflect the spatial variation patterns of regional rainfall, thus affecting the accuracy of reconstructing the spatial distribution of regional rainfall.

[0006] Therefore, there is currently a lack of a method that can improve the accuracy of regional rainfall spatial distribution reconstruction, reduce the cost of rain gauge construction and site selection, and provide technical and data support for extreme weather forecasting. Summary of the Invention

[0007] This application provides a method for reconstructing the spatial distribution of regional rainfall based on deep learning, in order to overcome the shortcomings of the aforementioned related technologies. The technical solution is as follows: Firstly, this application provides a method for reconstructing the spatial distribution of regional rainfall based on deep learning, including: The target area is gridded, and historical elevation data and rainfall observation data are filled into the corresponding grids to obtain the elevation and monthly rainfall of each grid. Calculate the slope and monthly rainfall variation gradient for each grid, and divide the feature sub-regions according to the numerical distribution of elevation, slope, monthly rainfall and monthly rainfall variation gradient for all grids, and determine the feature sub-region in which each grid falls. Rain gauge stations are randomly deployed within each feature sub-region, and corresponding feature masks are constructed based on their deployment locations. Rainfall reference data within a reference period are selected to perform normalization correction on the rainfall observation data, resulting in rainfall label data. The historical period is divided into a training period and a testing period. Rainfall label data and feature masks of each type of feature sub-region are selected during the training period to train the rainfall spatial distribution interpolation model, thus obtaining the trained rainfall spatial distribution interpolation model. Rainfall label data from the test period and feature masks for each type of feature sub-region were selected to test the trained rainfall spatial distribution interpolation model. Based on the test results, the feature sub-region with the highest accuracy was selected as the recommended deployment area for rain gauges.

[0008] In one alternative to the first aspect, calculating the slope and monthly rainfall variation gradient for each grid includes: The meridional slope of the current grid is calculated based on the elevation difference between the two grids immediately adjacent to it in the meridional direction, the latitudinal slope of the current grid is calculated based on the elevation difference between the two grids immediately adjacent to it in the latitudinal direction, and the slope of the current grid is calculated based on the meridional slope and the latitudinal slope. The meridional monthly rainfall variation gradient of the current grid is calculated based on the difference in monthly rainfall between the two grids immediately adjacent to it in the meridional direction. The zonal monthly rainfall variation gradient of the current grid is calculated based on the difference in monthly rainfall between the two grids immediately adjacent to it in the zonal direction. The monthly rainfall variation gradient of the current grid is calculated based on both the meridional and zonal monthly rainfall variation gradients.

[0009] In one alternative embodiment of the first aspect, the step of dividing the feature sub-regions according to the numerical distribution of elevation, slope, monthly rainfall, and monthly rainfall variation gradient of all grids, and determining the feature sub-region to which each grid falls, includes: The numerical distribution of all grids within the target area was statistically analyzed for four parameters: elevation, slope, monthly rainfall, and monthly rainfall variation gradient. Construct a frequency histogram based on the numerical distribution of each parameter type, and select the first and second quantiles for each parameter type based on the frequency histograms; wherein the first quantile is less than the second quantile. For each type of parameter, extract the value of the corresponding type parameter for each grid, and divide the grids with values ​​less than or equal to the first quantile, the grids with values ​​falling between the first and second quantiles, and the grids with values ​​greater than or equal to the second quantile into three different feature sub-regions respectively. Obtain all feature sub-regions corresponding to each parameter.

[0010] In one alternative embodiment of the first aspect, the step of randomly deploying rain gauges within each type of feature sub-region and constructing corresponding feature masks based on their deployment locations includes: Based on the preset rain gauge deployment density, the deployment location of rain gauges is randomly selected in each type of feature sub-region to determine the grid corresponding to the deployment location. Set the mask value of the grid corresponding to the deployment location to 1, and set the mask value of other grids in the target area to 0 to construct the feature mask of the current deployment scheme.

[0011] In one alternative embodiment of the first aspect, the step of normalizing the rainfall observation data using rainfall reference data within a selected reference period to obtain rainfall label data includes: Obtain rainfall reference data for the target area during the reference period, and calculate the grid average monthly rainfall for each grid in the target area during the reference period and the regional average monthly rainfall for the entire target area during the reference period based on the rainfall reference data. The rainfall observation data is normalized based on the grid average monthly rainfall and the regional average monthly rainfall. The difference between the monthly rainfall and the average monthly rainfall in each grid in the target area during the historical period is calculated. The ratio of the rainfall difference to the regional average monthly rainfall is used as the corrected monthly rainfall. The corrected monthly rainfall for each grid within the target area during the historical period is obtained as rainfall label data.

[0012] In one alternative embodiment of the first aspect, the step of selecting rainfall label data during the training period and feature masks for each type of feature sub-region to train the rainfall spatial distribution interpolation model, thereby obtaining the trained rainfall spatial distribution interpolation model, includes: Multiply the feature mask corresponding to each type of feature sub-region with the rainfall label data during the training period, and fill the extracted monthly rainfall with the observation data of each rain gauge station under the deployment scheme corresponding to the feature mask. The monthly rainfall data of grids other than the rain gauge grids is reconstructed by interpolating the rainfall spatial distribution state interpolation model based on the observation data, and the regional rainfall spatial distribution of the target area under the current feature mask layout scheme is obtained. A loss function is constructed based on the difference between the reconstructed regional rainfall spatial distribution and rainfall label data; Backpropagation is performed based on the value of the loss function to optimize the model parameters of the rainfall spatial distribution interpolation model; Determine the model parameters corresponding to the minimum value of the loss function to obtain the trained spatial distribution interpolation model of rainfall.

[0013] In one alternative of the first aspect, the step of selecting rainfall label data for the test period and feature masks for each type of feature sub-region to test the trained rainfall spatial distribution interpolation model, and selecting the feature sub-region with the highest accuracy as the recommended deployment area for rain gauges based on the test results, includes: Multiply the feature mask corresponding to each type of feature sub-region with the rainfall label data during the test period, and fill the extracted monthly rainfall into the test data of each rain gauge station under the deployment scheme corresponding to the feature mask. By interpolating the trained rainfall spatial distribution state interpolation model based on the test data, the monthly rainfall of other grids besides the rain gauge grid is reconstructed, and the predicted regional rainfall spatial distribution of the target area under the current feature mask layout scheme is obtained. For each type of feature sub-region, the determination coefficient corresponding to the feature sub-region is calculated based on the difference between the monthly rainfall of each grid in the spatial distribution of rainfall in the predicted area and the monthly rainfall of the corresponding grid in the rainfall label data. The determination coefficients corresponding to the feature sub-regions are used as the accuracy of the trained rainfall spatial distribution interpolation model in restoring the regional rainfall spatial distribution. Each type of feature sub-region is arranged according to the magnitude of the accuracy. The feature sub-region with the highest accuracy is selected as the recommended deployment area for rain gauges.

[0014] Secondly, this application also provides a deep learning-based device for reconstructing the spatial distribution of regional rainfall, comprising: The data acquisition unit performs gridding on the target area, filling historical elevation data and rainfall observation data into the corresponding grids to obtain the elevation and monthly rainfall of each grid. The feature region is divided into units. The slope and monthly rainfall variation gradient of each grid are calculated. Based on the numerical distribution of elevation, slope, monthly rainfall and monthly rainfall variation gradient of all grids, the feature sub-regions are divided, and the feature sub-region to which each grid falls is determined. The data processing unit is used to randomly deploy rain gauges in each type of feature sub-region and construct corresponding feature masks according to the deployment locations; the data processing unit is also used to select rainfall reference data within a reference period to perform normalization correction on the rainfall observation data to obtain rainfall label data. The model training unit is used to divide the historical period into a training period and a testing period, select the rainfall label data and feature masks of each type of feature sub-region during the training period to train the rainfall spatial distribution interpolation model, and obtain the trained rainfall spatial distribution interpolation model. The deployment area selection unit is used to select the rainfall label data during the test period and the feature mask of each type of feature sub-region to test the trained rainfall spatial distribution interpolation model. Based on the test results, the feature sub-region with the highest accuracy is selected as the recommended deployment area for the rain gauge.

[0015] Thirdly, this application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method provided by the first aspect of this application or any implementation thereof.

[0016] Fourthly, this application also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method provided by the first aspect of this application or any implementation thereof.

[0017] The beneficial effects of the technical solution provided in this application include at least the following: This application divides characteristic sub-regions and constructs characteristic masks by gridding elevation, slope, monthly rainfall and variation gradient, so that the rain gauges can hide regional topography and macro-rainfall characteristics. Under the same station density, it significantly improves the model's accuracy in reconstructing the spatial distribution of rainfall, thereby reducing the construction cost of the observation network.

[0018] This application utilizes training and testing data to train and optimize the neural network model. This not only enables the model to adaptively learn the representation ability of different feature sub-regions of rainfall spatial structure in actual interpolation, but also automatically selects the recommended deployment area with the highest accuracy. This provides a scientific theoretical basis for rain gauge site selection and avoids the problem of inefficient observation information caused by improper station location in traditional methods.

[0019] This application can more accurately reconstruct the spatial distribution of regional rainfall under limited rain gauge conditions, effectively supporting extreme weather forecasting and water resource assessment, while optimizing the station network layout and achieving a balance between improved accuracy and cost control. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in this application or related technologies, the drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart illustrating a method for restoring the spatial distribution of regional rainfall based on deep learning, as provided in an embodiment of this application. Figure 2 This is a schematic diagram of a deep learning-based device for reconstructing the spatial distribution of regional rainfall, as provided in an embodiment of this application. Figure 3 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0023] The terms "comprising" and "having," and any variations thereof, in the specification, claims, and accompanying drawings of this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or modules is not limited to the steps or modules listed, but may optionally include steps or modules not listed, or may optionally include other steps or modules inherent to such process, method, product, or apparatus.

[0024] It should be noted that the terms "first" and "second" used in this application are merely to distinguish similar objects and do not represent a specific ordering of the objects. It is understood that "first" and "second" can be interchanged in a specific order or sequence where permitted. It should be understood that the objects distinguished by "first" and "second" can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in an order other than those described or illustrated herein.

[0025] The present application will now be described in detail with reference to specific embodiments.

[0026] Next, combine Figure 1 This application introduces a method for reconstructing the spatial distribution of regional rainfall based on deep learning, as provided in its embodiments. For details, please refer to... Figure 1 , Figure 1 This illustration shows a flowchart of a method for reconstructing the spatial distribution of regional rainfall based on deep learning, as provided in an embodiment of this application. Figure 1 As shown, the method includes the following steps: S101, the target area is gridded, and historical elevation data and rainfall observation data are filled into the corresponding grid to obtain the elevation and monthly rainfall of each grid; S102, calculate the slope and monthly rainfall variation gradient of each grid, divide the feature sub-regions according to the numerical distribution of elevation, slope, monthly rainfall and monthly rainfall variation gradient of all grids, and determine the feature sub-region in which each grid falls. S103, Rain gauge stations are randomly deployed in each type of feature sub-region, and corresponding feature masks are constructed according to the deployment locations; Rainfall reference data within the reference period are selected to perform normalization correction on the rainfall observation data to obtain rainfall label data; S104, the historical period is divided into a training period and a testing period. The rainfall label data and feature masks of each type of feature sub-region are selected during the training period to train the rainfall spatial distribution interpolation model, and the trained rainfall spatial distribution interpolation model is obtained. S105. Select the rainfall label data of the test period and the feature mask of each type of feature sub-region to test the trained rainfall spatial distribution interpolation model. Based on the test results, select the feature sub-region with the highest accuracy as the recommended deployment area of ​​the rain gauge.

[0027] In some embodiments, in S101, any province, city, or municipal district can be selected as the target area for study, and the target area can be divided into multiple latitude and longitude grids of the same size based on a preset spatial resolution and at equal latitude and longitude intervals.

[0028] Specifically, elevation data and rainfall observation data from any historical period can be selected as the basis for subsequent model training and testing. For example, in this embodiment, a monthly-scale data sequence from 1950 to 2014 is selected, containing rainfall observation values ​​for each month within the historical period. Elevation data of the target area can be obtained from digital elevation models, topographic maps, and other materials. Monthly-scale rainfall observation data of the target area at different spatial resolutions within the historical period can be collected from global rainfall datasets from different sources.

[0029] Understandably, the elevation and monthly rainfall of the corresponding coordinates can be selected from the elevation data and rainfall observation data based on the grid coordinates, thereby obtaining the elevation and monthly rainfall of each grid.

[0030] In some embodiments, S102, the slope and monthly rainfall variation gradient for each grid can be calculated based on the following steps: The meridional slope of the current grid is calculated based on the elevation difference between the two grids immediately adjacent to it in the meridional direction. The formula is as follows: ; The latitudinal slope of the current grid is calculated based on the elevation difference between the two grids immediately adjacent to it in the latitudinal direction. The formula is as follows: ; The slope of the current grid is calculated based on the meridional and zonal slopes, expressed by the following formula: ; The gradient of monthly rainfall variation in the meridional direction of the current grid is calculated based on the difference in monthly rainfall between the two grids immediately adjacent to it in the meridional direction. The formula is as follows: ; The zonal monthly rainfall gradient of the current grid is calculated based on the difference in monthly rainfall between the two grids immediately adjacent to it in the zonal direction. The formula is as follows: ; The combined monthly rainfall gradient for the current grid is calculated based on the meridional and zonal monthly rainfall gradients, expressed by the following formula: ; in, This represents the current grid, where i represents the latitude label of the current grid and j represents the longitude label of the current grid. , This represents the two grids that are immediately adjacent to the current grid in the longitudinal direction. , This represents the two grids that are immediately adjacent to the current grid in the latitudinal direction, and represents the current grid. Indicates the latitudinal spacing of the grid. Indicates the longitude spacing of the grid; The values ​​represent the grid elevation of the corresponding grid; grad_lat_ele represents the latitudinal slope of the corresponding grid, grad_lon_ele represents the meridional slope of the corresponding grid, and gradient_ele represents the composite slope of the corresponding grid, which is used as the slope of the corresponding grid in subsequent calculations; precipitation represents the monthly precipitation of the corresponding grid; grad_lat_pre represents the latitudinal monthly precipitation variation gradient of the corresponding grid, grad_lon_pre represents the meridional monthly precipitation variation gradient of the corresponding grid, and gradient_pre represents the composite monthly precipitation variation gradient of the corresponding grid, which is used as the monthly precipitation variation gradient of the corresponding grid in subsequent calculations.

[0031] Furthermore, based on the slope and monthly rainfall variation gradient of each grid obtained above, S102 can perform subsequent steps to divide the feature sub-regions, including the following steps: S201, respectively statistically analyzes the numerical distribution of all grids within the target area on four types of parameters: elevation, slope, monthly rainfall, and monthly rainfall variation gradient.

[0032] S202, construct the corresponding frequency histogram based on the numerical distribution of each type of parameter, and select the first quantile and the second quantile corresponding to each type of parameter based on the frequency histogram.

[0033] Specifically, the process of constructing a frequency histogram can be summarized as follows: for each type of parameter, the total range from the minimum to the maximum value in the numerical distribution can be divided into several numerical intervals. The frequency of the corresponding parameter falling into the corresponding numerical interval is obtained by taking the horizontal axis as the numerical interval and the vertical axis as the grid. The number of grids falling into each numerical interval can be observed intuitively from the frequency histogram.

[0034] The first and second quantiles can be set according to the actual situation. For example, the parameter value at the 30th percentile can be set as the first quantile and the parameter value at the 70th percentile. The corresponding 30th percentile and 70th percentile can be selected according to the numerical distribution of the four types of parameters. The specific values ​​of the quantiles are not limited in the embodiments of this application.

[0035] For example, taking elevation as an example, the total range of all grids in elevation is 0-100m. The 30th quantile and the 70th quantile can be selected. The elevation value corresponding to the 30th quantile is 30m, and the elevation value corresponding to the 70th quantile is 100m.

[0036] S203, for each type of parameter, extract the value of the corresponding type parameter for each grid, and divide the grids with values ​​less than or equal to the first quantile, the grids with values ​​falling between the first quantile and the second quantile, and the grids with values ​​greater than or equal to the second quantile into three different feature sub-regions.

[0037] Specifically, the corresponding value of each grid can be compared with the first and second quantiles. Taking elevation as an example, if the elevation of a grid is 10m, and 10m is less than the first quantile, then the grid is classified into the characteristic sub-region of less than or equal to 30m, i.e., the low elevation region; if the elevation of a grid is 45m, and 45m is within the interval (30m, 70m), then the grid is classified into the characteristic sub-region of the interval (30m, 70m), i.e., the medium elevation region.

[0038] For example, each type of parameter can be divided into three levels according to the above classification method: those less than or equal to the first quantile are low-level, those greater than or equal to the second quantile are high-level, and those falling within the range of the first and second quantiles are mid-level.

[0039] S204, obtains all feature sub-regions corresponding to each parameter.

[0040] Specifically, this application provides four types of parameters, resulting in a total of 12 types of feature sub-regions, including high elevation region, medium elevation region, low elevation region, high slope region, medium slope region, low slope region, high monthly rainfall region, medium monthly rainfall region, low monthly rainfall region, high monthly rainfall variation gradient region, medium monthly rainfall variation gradient region, and low monthly rainfall variation gradient region.

[0041] Furthermore, feature masks can be constructed separately for each type of feature sub-region, S103 including: Based on the preset rain gauge deployment density, the deployment location of rain gauges is randomly selected in each type of feature sub-region to determine the grid corresponding to the deployment location. Set the mask value of the grid corresponding to the deployment location to 1, and set the mask value of other grids in the target area to 0 to construct the feature mask of the current deployment scheme.

[0042] Understandably, the density of rain gauges reflects the number of rain gauges per unit area. One rain gauge can provide the rainfall observation value at a grid location. Therefore, the density of rain gauges is equivalent to the density of rainfall observation values. The specific value can depend on the accuracy of the model. The higher the accuracy and the better the deployment conditions in the area, the higher the density of rainfall observation values ​​can be. This application does not limit this.

[0043] In some embodiments, the normalization correction step in S103 can be performed simultaneously with the feature mask construction process. Alternatively, the normalization correction can be performed after extracting historical rainfall observation data in S101 and then filling it into the grid. Or, the normalization correction can be performed after filling the historical rainfall observation data into the grid. This application embodiment does not limit this.

[0044] The specific steps of normalization correction include: S301, Obtain rainfall reference data for the target area during the reference period, and calculate the grid average monthly rainfall for each grid within the target area during the reference period and the regional average monthly rainfall for the entire target area during the reference period based on the rainfall reference data.

[0045] Specifically, the reference period can be selected as needed from a period of climate stability within the historical timeframe of the target region. A climate-stable period can be understood as a period in which the global and regional climate systems are relatively stable overall, the trends of climate element changes are weak, and there are usually continuous and complete observation records. For example, 1961-1990 can be selected as the reference period, but this application does not limit this.

[0046] In this embodiment of the application, the historical period can be selected as 1950 to 2014, the data from 1961 to 1990 can be classified as the reference period, the data from 1950 to 2000 can be classified as the training period, and the data from 2001 to 2014 can be classified as the test period.

[0047] S302, based on the grid-average monthly rainfall and the regional average monthly rainfall, normalize the rainfall observation data. Calculate the difference between the monthly rainfall and the average monthly rainfall for each grid within the target area during the historical period. The ratio of this difference to the regional average monthly rainfall is used as the corrected monthly rainfall. The formula is as follows: ; Among them, subscript Indicates the corresponding month number. Indicates the corresponding longitude designation. Indicates the corresponding latitude label; This represents the monthly rainfall for month t at a grid corresponding to longitude and latitude. This represents the corrected monthly rainfall for the corresponding longitude and latitude grid in month t; This represents the average monthly rainfall of the grid corresponding to the longitude and latitude within the target area during the reference period; This represents the regional average monthly rainfall for the entire target area during the reference period.

[0048] S303, obtain the corrected monthly rainfall for each grid within the target area for each month in the historical period as rainfall label data.

[0049] Specifically, the normalized monthly rainfall can be considered as the rainfall under the true historical conditions. The corrected monthly rainfall for each grid in the historical period can be used as rainfall label data for subsequent model training, serving as the ground truth label that the model needs to learn.

[0050] In some embodiments, the rainfall spatial distribution state interpolation model provided in this application can be built based on a partially convolutional U-Net network framework, including an input layer, three encoder modules, three decoder modules, and an output layer.

[0051] Each encoder module consists of a partially convolutional module, which sequentially includes 2D partially convolution, batch normalization, and a ReLU activation function. The encoder modules are stacked hierarchically; as the number of layers increases, the spatial resolution of the feature maps gradually decreases while the number of channels increases accordingly. Skip connections are used to pass the encoded features of the corresponding layer to the decoding stage. Each decoder module also consists of a partially convolutional module. An upsampling layer is introduced before the convolution operation to restore the spatial resolution, and then the features are fused with the corresponding features from the encoder layer, sequentially passing through partially convolution, batch normalization, and a LeakyReLU activation function. The output layer maps the final features of the decoder to the predicted spatial distribution of regional rainfall. The overall network adopts a symmetrical encoder-decoder structure and incorporates skip connections to achieve multi-scale feature fusion.

[0052] The training process of the model is described below. S104 includes the following steps: S1041, multiply the feature mask corresponding to each type of feature sub-region with the rainfall label data during the training period, and fill the extracted monthly rainfall with the observation data of each rain gauge station under the deployment scheme corresponding to the feature mask.

[0053] Understandably, the monthly rainfall at the corresponding grid point can be extracted from the rainfall label data by using the grid points with a value of 1 in the feature mask. The extracted data is considered to be the observation data at the corresponding grid point of the rain gauge under the assumed rain gauge deployment scheme. On the other hand, the grid points with a value of 0 in the feature mask are considered to be grid points with unknown rainfall that need to be filled by interpolation calculation.

[0054] S1042 uses a rainfall spatial distribution interpolation model to interpolate based on observation data, reconstructs the monthly rainfall of grids other than rain gauge grids, and obtains the regional rainfall spatial distribution of the target area under the current feature mask layout scheme.

[0055] S1043, construct a loss function based on the difference between the reconstructed regional rainfall spatial distribution and rainfall label data.

[0056] Specifically, the rainfall spatial distribution interpolation model can interpolate based on the observation data at the corresponding grid positions of the known rain gauge stations. The rainfall amount of each grid in the constructed regional rainfall spatial distribution is compared with the ground truth label in the rainfall label data corresponding to the same latitude and longitude grid. The reconstruction error is constructed based on the difference between the two, and the loss function is constructed based on the reconstruction error.

[0057] S1044, backpropagation is performed based on the value of the loss function to optimize the model parameters of the rainfall spatial distribution interpolation model.

[0058] For example, the Adam optimizer can be used to update model parameters during model training, with a learning rate set to 2×10. -4 The maximum number of iterations is set to 500,000 to achieve iterative optimization of model parameters, but this application does not limit this.

[0059] S1045, determine the model parameters corresponding to the minimum value of the loss function, and obtain the trained spatial distribution interpolation model of rainfall.

[0060] It should be noted that, in this embodiment, feature sub-regions are divided and feature masks are constructed by gridding elevation, slope, monthly rainfall and gradient changes. This allows rain gauges to implicitly represent regional topography and macro-rainfall characteristics. Based on this, the rain gauges are optimized and their locations are reasonably allocated. This enables the model to implicitly learn regional topography information and macro-rainfall characteristics during training. Consequently, more representative observation information on the spatial distribution of regional rainfall can be obtained during actual interpolation, providing a foundation for more accurately reconstructing the spatial distribution of regional rainfall.

[0061] Next, the rainfall label data for the test period and the feature masks of each type of feature sub-region can be selected to test the trained rainfall spatial distribution interpolation model. Based on the test results, the feature sub-region with the highest accuracy is selected as the recommended deployment area for rain gauges. S105 includes: S501, multiply the feature mask corresponding to each type of feature sub-region with the rainfall label data during the test period, and fill the extracted values ​​into the test data of each rain gauge station under the deployment scheme corresponding to the feature mask.

[0062] Specifically, similar to S401 during model training, during testing, feature masks can be applied to the rainfall label data during the testing period to obtain the test data of the corresponding feature sub-regions during the testing period.

[0063] S502 uses a trained rainfall spatial distribution interpolation model to perform interpolation processing based on test data, reconstructs the monthly rainfall of grids other than rain gauge grids, and obtains the predicted regional rainfall spatial distribution of the target area under the current feature mask layout scheme.

[0064] S503, for each type of feature sub-region, the determination coefficient corresponding to the feature sub-region is calculated based on the difference between the monthly rainfall of each grid in the predicted area's rainfall spatial distribution and the monthly rainfall of the corresponding grid in the rainfall label data. The formula is expressed as: ; in, This represents the monthly rainfall in the k-th grid of the rainfall label data. This represents the monthly rainfall in the k-th grid of the predicted regional rainfall spatial distribution. This represents the average rainfall across all grid cells in the rainfall label data.

[0065] It should be noted that during testing, the coefficient of determination is calculated for each predicted area rainfall spatial distribution obtained by processing all feature masks of each type of feature sub-region. The calculated coefficient of determination reflects the accuracy of the trained model in reconstructing the spatial distribution of regional rainfall under the rain gauge deployment scheme of this type of feature sub-region.

[0066] S504, the determination coefficients corresponding to the feature sub-regions are used as the accuracy of the trained rainfall spatial distribution interpolation model in restoring the regional rainfall spatial distribution, and each type of feature sub-region is arranged according to the numerical value of the accuracy.

[0067] S505 uses the feature sub-region with the highest accuracy as the recommended deployment area for rain gauges.

[0068] In this embodiment of the application, target regions A, B, C, D, E, and F are used as target areas to implement a deep learning-based method for restoring the spatial distribution of regional rainfall provided in this embodiment of the application. The accuracy of the trained rainfall spatial distribution interpolation model in restoring the spatial distribution of regional rainfall is tested when applied to these regions. The results of the model provided in this embodiment of the application are compared with those of the inverse distance weighted interpolation method and the Kriging interpolation method, as shown in Table 1.

[0069] Table 1. Accuracy of restoring the spatial distribution of rainfall in the region

[0070] As shown in Table 1, compared with the inverse distance weighted interpolation method and the Kriging interpolation method, the scheme provided by the embodiments of this application achieves higher accuracy. At the same time, the rainfall spatial distribution state interpolation model trained in S104 of the embodiments of this application can determine the most suitable feature sub-region for building rain gauge stations in the corresponding target area based on the accuracy value. For example, target area A achieves the highest accuracy under the rain gauge station layout scheme of the feature sub-region "high monthly rainfall variation gradient area", and target area B achieves the highest accuracy under the rain gauge station layout scheme of the feature sub-region "high monthly rainfall variation gradient area".

[0071] In this way, the embodiments of this application can not only restore the spatial distribution of regional rainfall within the target area based on the trained model, but also determine the recommended deployment areas of rain gauges through the trained model, thereby providing a theoretical basis for the site selection of rain gauges by relevant departments / institutions, providing better site selection for the deployment of rain gauges, and the optimized rain gauge deployment scheme can better characterize the spatial structure of rainfall in the target area, which is conducive to improving the reconstruction accuracy of the spatial distribution of regional rainfall.

[0072] The following are apparatus embodiments of this application, which can be used to execute the method embodiments of this application. For details not disclosed in the apparatus embodiments of this application, please refer to the method embodiments of this application.

[0073] Please see below. Figure 2 The image below is a schematic diagram of a deep learning-based device for reconstructing the spatial distribution of regional rainfall, provided as an exemplary embodiment of this application. The device includes: The data acquisition unit performs gridding on the target area, filling historical elevation data and rainfall observation data into the corresponding grids to obtain the elevation and monthly rainfall of each grid. The feature region is divided into units. The slope and monthly rainfall variation gradient of each grid are calculated. Based on the numerical distribution of elevation, slope, monthly rainfall and monthly rainfall variation gradient of all grids, the feature sub-regions are divided, and the feature sub-region to which each grid falls is determined. The data processing unit is used to randomly deploy rain gauges in each type of feature sub-region and construct corresponding feature masks according to the deployment locations; the data processing unit is also used to select rainfall reference data within a reference period to perform normalization correction on the rainfall observation data to obtain rainfall label data. The model training unit is used to divide the historical period into a training period and a testing period, select the rainfall label data and feature masks of each type of feature sub-region during the training period to train the rainfall spatial distribution interpolation model, and obtain the trained rainfall spatial distribution interpolation model. The deployment area selection unit is used to select the rainfall label data during the test period and the feature mask of each type of feature sub-region to test the trained rainfall spatial distribution interpolation model. Based on the test results, the feature sub-region with the highest accuracy is selected as the recommended deployment area for the rain gauge.

[0074] It should be noted that the above embodiments, when implementing a deep learning-based method for restoring the spatial distribution of regional rainfall, are only illustrative examples of the division of the functional modules described above. In practical applications, the functions described above can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. Furthermore, the device and method embodiments provided above belong to the same concept, and their implementation process is detailed in the method embodiments, which will not be repeated here.

[0075] This application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of any of the methods described above.

[0076] Please see Figure 3 This is a structural block diagram of an electronic device provided in an embodiment of this application.

[0077] like Figure 3 As shown, the electronic device includes a processor and a memory.

[0078] In this embodiment, the processor is the control center of the computer system, and can be a processor of a physical machine or a processor of a virtual machine. The processor may include one or more processing cores, such as a 4-core processor, an 8-core processor, etc. The processor can be implemented using at least one hardware form of DSP (Digital Signal Processing), FPGA (Field-Programmable Gate Array), or PLA (Programmable Logic Array).

[0079] A processor can also include a main processor and a coprocessor. The main processor is used to process data in the wake-up state and is also called the CPU (Central Processing Unit). The coprocessor is a low-power processor used to process data in the standby state.

[0080] The memory may include one or more computer-readable storage media, which may be non-transitory. The memory may also include high-speed random access memory and non-volatile memory, such as one or more disk storage devices or flash memory devices. In some embodiments of this application, the non-transitory computer-readable storage media in the memory are used to store at least one instruction, which is executed by a processor to implement the methods in the embodiments of this application.

[0081] In some embodiments, the electronic device further includes a peripheral device interface and at least one peripheral device. The processor, memory, and peripheral device interface are connected via a bus or signal line. Each peripheral device is connected to the peripheral device interface via a bus, signal line, or circuit board. Specifically, the peripheral device includes: a display screen, a camera, and audio circuitry. The peripheral device interface can be used to connect at least one I / O (Input / Output) related peripheral device to the processor and memory.

[0082] In some embodiments of this application, the processor, memory, and peripheral device interfaces are integrated on the same chip or circuit board; in other embodiments of this application, any one or two of the processor, memory, and peripheral device interfaces can be implemented on separate chips or circuit boards. This application does not specifically limit the implementation in this regard.

[0083] The electronic device structural block diagrams shown in the embodiments of this application do not constitute a limitation on the electronic device. The electronic device may include more or fewer components than shown, or combine certain components, or use different component arrangements.

[0084] This application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the methods in any of the foregoing embodiments. The computer-readable storage medium may include, but is not limited to, any type of disk, including floppy disks, optical disks, DVDs, CD-ROMs, microdrives, as well as magneto-optical disks, ROMs, RAMs, EPROMs, EEPROMs, DRAMs, VRAMs, flash memory devices, magnetic cards or optical cards, nanosystems (including molecular memory ICs), or any type of medium or device suitable for storing instructions and / or data.

[0085] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the parts that contribute to the related technology, can be embodied in the form of software products. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0086] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for reconstructing the spatial distribution of regional rainfall based on deep learning, characterized in that, include: The target area is gridded, and historical elevation data and rainfall observation data are filled into the corresponding grids to obtain the elevation and monthly rainfall of each grid. Calculate the slope and monthly rainfall variation gradient for each grid, and divide the feature sub-regions according to the numerical distribution of elevation, slope, monthly rainfall and monthly rainfall variation gradient for all grids, and determine the feature sub-region in which each grid falls. Rain gauges are randomly deployed within each type of feature sub-region, and corresponding feature masks are constructed based on their deployment locations. Normalization correction was performed on the rainfall observation data by selecting rainfall reference data within the reference period to obtain rainfall label data; The historical period is divided into a training period and a testing period. Rainfall label data and feature masks of each type of feature sub-region are selected during the training period to train the rainfall spatial distribution interpolation model, thus obtaining the trained rainfall spatial distribution interpolation model. Rainfall label data from the test period and feature masks for each type of feature sub-region were selected to test the trained rainfall spatial distribution interpolation model. Based on the test results, the feature sub-region with the highest accuracy was selected as the recommended deployment area for rain gauges, specifically including: Multiply the feature mask corresponding to each type of feature sub-region with the rainfall label data during the test period, and fill the extracted monthly rainfall into the test data of each rain gauge station under the deployment scheme corresponding to the feature mask. By interpolating the trained rainfall spatial distribution state interpolation model based on the test data, the monthly rainfall of other grids besides the rain gauge grid is reconstructed, and the predicted regional rainfall spatial distribution of the target area under the current feature mask layout scheme is obtained. For each type of feature sub-region, the determination coefficient corresponding to the feature sub-region is calculated based on the difference between the monthly rainfall of each grid in the spatial distribution of rainfall in the predicted area and the monthly rainfall of the corresponding grid in the rainfall label data. The determination coefficients corresponding to the feature sub-regions are used as the accuracy of the trained rainfall spatial distribution interpolation model in restoring the regional rainfall spatial distribution. Each type of feature sub-region is arranged according to the magnitude of the accuracy. The feature sub-region with the highest accuracy is selected as the recommended deployment area for rain gauges.

2. The method for reconstructing the spatial distribution of regional rainfall based on deep learning according to claim 1, characterized in that, The calculation of the slope and monthly rainfall variation gradient for each grid includes: The meridional slope of the current grid is calculated based on the elevation difference between the two grids immediately adjacent to it in the meridional direction, the latitudinal slope of the current grid is calculated based on the elevation difference between the two grids immediately adjacent to it in the latitudinal direction, and the slope of the current grid is calculated based on the meridional slope and the latitudinal slope. The meridional monthly rainfall variation gradient of the current grid is calculated based on the difference in monthly rainfall between the two grids immediately adjacent to it in the meridional direction. The zonal monthly rainfall variation gradient of the current grid is calculated based on the difference in monthly rainfall between the two grids immediately adjacent to it in the zonal direction. The monthly rainfall variation gradient of the current grid is calculated based on both the meridional and zonal monthly rainfall variation gradients.

3. The method for reconstructing the spatial distribution of regional rainfall based on deep learning according to claim 2, characterized in that, The process involves dividing the grid into feature sub-regions based on the numerical distribution of elevation, slope, monthly rainfall, and monthly rainfall variation gradient across all grids, and determining the feature sub-region to which each grid falls, including: The numerical distribution of all grids within the target area was statistically analyzed for four parameters: elevation, slope, monthly rainfall, and monthly rainfall variation gradient. Construct a frequency histogram based on the numerical distribution of each parameter type, and select the first and second quantiles for each parameter type based on the frequency histograms; wherein the first quantile is less than the second quantile. For each type of parameter, extract the value of the corresponding type parameter for each grid, and divide the grids with values ​​less than or equal to the first quantile, the grids with values ​​falling between the first and second quantiles, and the grids with values ​​greater than or equal to the second quantile into three different feature sub-regions respectively. Obtain all feature sub-regions corresponding to each parameter.

4. The method for reconstructing the spatial distribution of regional rainfall based on deep learning according to claim 1, characterized in that, The step of randomly deploying rain gauges within each type of feature sub-region and constructing corresponding feature masks based on their deployment locations includes: Based on the preset rain gauge deployment density, the deployment location of rain gauges is randomly selected in each type of feature sub-region to determine the grid corresponding to the deployment location. Set the mask value of the grid corresponding to the deployment location to 1, and set the mask value of other grids in the target area to 0 to construct the feature mask of the current deployment scheme.

5. The method for reconstructing the spatial distribution of regional rainfall based on deep learning according to claim 1, characterized in that, The rainfall observation data is normalized by selecting rainfall reference data within a reference period to obtain rainfall label data, including: Obtain rainfall reference data for the target area during the reference period, and calculate the grid average monthly rainfall for each grid in the target area during the reference period and the regional average monthly rainfall for the entire target area during the reference period based on the rainfall reference data. The rainfall observation data is normalized based on the grid average monthly rainfall and the regional average monthly rainfall. The difference between the monthly rainfall and the average monthly rainfall in each grid in the target area during the historical period is calculated. The ratio of the rainfall difference to the regional average monthly rainfall is used as the corrected monthly rainfall. The corrected monthly rainfall for each grid within the target area during the historical period is obtained as rainfall label data.

6. The method for reconstructing the spatial distribution of regional rainfall based on deep learning according to claim 4, characterized in that, The selection of rainfall label data during the training period and feature masks for each type of feature sub-region are used to train the rainfall spatial distribution interpolation model, resulting in a trained rainfall spatial distribution interpolation model, including: Multiply the feature mask corresponding to each type of feature sub-region with the rainfall label data during the training period, and fill the extracted monthly rainfall with the observation data of each rain gauge station under the deployment scheme corresponding to the feature mask. The monthly rainfall data of grids other than the rain gauge grids is reconstructed by interpolating the rainfall spatial distribution state interpolation model based on the observation data, and the regional rainfall spatial distribution of the target area under the current feature mask layout scheme is obtained. A loss function is constructed based on the difference between the reconstructed regional rainfall spatial distribution and rainfall label data; Backpropagation is performed based on the value of the loss function to optimize the model parameters of the rainfall spatial distribution interpolation model; Determine the model parameters corresponding to the minimum value of the loss function to obtain the trained spatial distribution interpolation model of rainfall.

7. An apparatus for reconstructing the spatial distribution of regional rainfall based on deep learning as described in any one of claims 1-6, characterized in that, The device includes: The data acquisition unit performs gridding on the target area, filling historical elevation data and rainfall observation data into the corresponding grids to obtain the elevation and monthly rainfall of each grid. The feature region is divided into units. The slope and monthly rainfall variation gradient of each grid are calculated. Based on the numerical distribution of elevation, slope, monthly rainfall and monthly rainfall variation gradient of all grids, the feature sub-regions are divided, and the feature sub-region to which each grid falls is determined. The data processing unit is used to randomly deploy rain gauges in each type of feature sub-region and construct corresponding feature masks according to the deployment locations; the data processing unit is also used to select rainfall reference data within a reference period to perform normalization correction on the rainfall observation data to obtain rainfall label data. The model training unit is used to divide the historical period into a training period and a testing period, select the rainfall label data and feature masks of each type of feature sub-region during the training period to train the rainfall spatial distribution interpolation model, and obtain the trained rainfall spatial distribution interpolation model. The deployment area selection unit is used to select the rainfall label data during the test period and the feature mask of each type of feature sub-region to test the trained rainfall spatial distribution interpolation model. Based on the test results, the feature sub-region with the highest accuracy is selected as the recommended deployment area for the rain gauge.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the deep learning-based regional rainfall spatial distribution state restoration method as described in any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the deep learning-based method for restoring the spatial distribution of regional rainfall as described in any one of claims 1 to 6.

Images